A New Method for Quantitative Estimation of Rolling Bearings Under Variable Working Conditions

Quantitative estimation of fault severity in rolling bearings is crucial for making proper maintenance decisions. However, bearings usually operate under variable working conditions in actual industrial applications, which significantly affects the extraction of fault severity information. For addressing this issue, a new health indicator construction method based on a quantitative estimation neural network is proposed. First, a dataset is constructed, where each sample datum corresponds to two labels: a severity label and a speed label. The former represents the fault severity, while the latter denotes the running speed. Then, the quantitative estimation neural network is designed to extract speed-invariant and severity-variant features from the sample data. This network consists of three modules: a feature extractor that extracts features from the sample data, a speed discriminator that ensures the speed-invariance of the features, and a severity classifier that ensures the severity-variance of the features. Finally, relative similarity is used to measure health indicators for rolling bearings. Two bearing datasets collected under variable speeds are used to validate the effectiveness of the proposed method. The results demonstrate that the proposed method can effectively reduce the impact of variable speeds and quantitatively estimate the fault severity in rolling bearings.